1

Prediction and planning

Networks and people are good at prediction

Temporal difference learning

Word segmentation

Sentence comprehension

Forward modeling

Learn a model of the world

Use model to predict the consequences of actions

Select actions based on the best predicted outcome

Can we use our ability to predict to aid in selecting the

best response?

2

3

4

What is cognitive control?

The models we’ve looked at are largely “recognition-based”

(map input to corresponding output)

These models ignore the human ability to control which

response we produce (if we produce one at all)

Networks can’t learn to produce different responses to the same input

Animals have trouble doing this, too

Can PDP models account for

the flexibility of human behavior?

5

What is cognitive control?

We can override prepotent responses: You’re hungry. There’s a sandwich on your roommate’s desk, but you

don’t eat it.

We can ignore things in the environment that aren’t

relevant for the task at hand: You’re at a busy train station looking for a friend wearing a red coat, and

you only look at the faces of people who are wearing red. (Miller &

Cohen, 2001)

The networks we’ve looked at process all input equally.

We can perform multiple tasks at the same time: You’re writing an e-mail while listening to someone on the telephone.

Networks typically perform one tasks at a time.

6

What is cognitive control?

What must cognitive control entail?

Select appropriate perceptual information for processing (e.g.

only people wearing red)

Inhibit inappropriate responses (e.g. don’t eat that sandwich)

Maintain relevant contextual information (e.g. this friend likes

cream in his tea)

Most of the networks we’ve seen can’t do this.

Is cognitive control qualitatively different from other kinds of

knowledge or processes?

Example: Stroop task

15

The Stroop task

No effect of ink color on word

reading

When the color name conflicts

with the word, reaction times

are the slowest

Color naming is slower than

word reading

16

Automatic vs. controlled processes

Word reading automatic

Color naming controlled

When outputs conflict,

controlled process will be

slowed

Automatic: fast, don’t require attention for execution, can

occur involuntarily

Controlled: slower, voluntary, require attention

17

MacLeod & Dunbar (1988)

Is there really a dichotomy between automatic and

controlled processes?

Taught subjects to use color names as names for neutral-colored

shapes

Initially, color naming interfered with shape naming

With extended training on shape naming, effects reversed

Speed of processing and interference depend on the degree of

automatization (due largely to practice)

Graded nature of effects suitable for connectionist modelling

“green”

18

PDP model of Stroop task (Cohen et al., 1990)

Separate pathways for

word reading and color

naming; Word reading

pathway is stronger

More practice, more

systematic task; doesn’t

need top-down support

Presence of a conflicting

color produces no

interference

Color naming requires

top-down support (control)

to override “prepotent”

response from word

pathway

19

At rest (R), a change in the net

input has little effect on

activation

After modulation by task units

(C), a change in the net input

has a larger impact on activation

Task information sensitizes

these units to external input

All units in pathway activated

equally

No specific information about

the correct response

PDP model of Stroop task (Cohen et al., 1990)

20

PDP model of Stroop task (Cohen et al., 1990)

Task demand units bias

processing in favor of the

weaker pathway

These units “guide” (or

implement) attention to

overcome the dominant

response

21

PDP model of Stroop task (Cohen et al., 1990)

22

Automaticity is a continuum of strength of processing

No qualitative distinction between “controlled” vs. “automatic”

processes

Same kinds of processing and representation used for

word reading and color naming participate in “cognitive

control”

PDP model of Stroop task (Cohen et al., 1990)

23

Attention & cognitive control

In the Stroop case, the task demand units are guiding

attention to enable cognitive control

Attention is:

“… the modulatory influence that representations of one type

have on selecting which (or to what degree) representations of

other types are processed…” (Cohen et al., 2004)

Attention biases competition between representations

competing to generate response

Bias can be “bottom-up” or “top-down”

In the Stroop case, it’s top-down (instructions given by the

experimenter to color name)

24

Cognitive control and the PFC

(Cohen et al., 2004)

Prefrontal cortex (PFC) subserves the function of the “task demand” units Can sustain the activation of

representations that “bias the flow of activity along task relevant pathways”

Models of PFC use recurrent connections “Attractor dynamics”

Units with mutually excitatory connections can actively maintain themselves without external input

25

Cognitive control and the PFC

(Cohen et al., 2004)

Lesions to parts of the PFC can produce deficits in “working memory” Patients are unable to maintain

task relevant information

Patients are also easily distracted during a task

26

PFC must also be able to update task representations

New input may signal a change in task or need to be ignored

The current degree of control may be insufficient to do the task

PFC needs to increase amount of “biasing”

Patients with lesions to the PFC may perseverate

Continuing to produce a response even when inappropriate or not

relevant to the task

How does the PFC know when to alter the current

amount of control?

Cognitive control and the PFC (Cohen et al., 2004)

27

Attention reduces conflicts in

processing

Occurrence of conflict signals

need for more attentional control

Anterior cingulate cortex (ACC)

appears to respond to conflict in

processing pathways and/or

response representations

Lesions to the ACC result in an

inability to detect errors and

severe difficulty with Stroop task

Cognitive control and the ACC (Cohen et al., 2004)

28

Cognitive control and the ACC

(Cohen et al., 2004)

29

Cognitive control & the VTA (Cohen et al., 2004)

How is a task representation

selected from many possible

representations?

Maybe via temporal difference

learning

Ventral Tegmental Area (VTA)

responds to errors in predicted

reward (Shulz et al., 1997) and

communicates with PFC

If a response is associated with

reward, ensure that relevant

external cues signal a task change

to the PFC in the future

30

Cognitive control: Summary (Cohen et al., 2004)

31

Cognitive control & expertise

Strong cognitive control (e.g. color naming in the Stroop

task) is effortful and errorful

We can do it, but it’s not easy

How can we reduce the cognitive demands of a difficult

task?

Novice chess players must study the board at length before

selecting a move, but experts “see” the move immediately

Learn to place more of the burden on the “recognition”

side of things

Humans and networks are inherently good at pattern recognition

32

Cognitive control & expertise

Experts encode perceptual input differently than novices

Chase & Simon (1975): participants viewed a chessboard with

pieces on it for 5 seconds

Later, they were asked to recall the positions of the pieces

Experts were better at recall when the pieces were in legal

configurations

No difference between experts and novices with the pieces were

placed randomly on the board

33

Cognitive control & expertise

The experts had become better at recognizing the

relevant aspects of the input

They could attend to the important details rather than all

information available

Another example: face processing

Experience guides attention, reducing the required

degree of top-down cognitive control

e.g. word reading in the Stroop task required less activation from

the task units

34

Summary

“Control processes” need not be different than any other

kinds of processing and representation

Amount of control depends on strength of the relevant task

Networks (and humans) can be more flexible than simply

“input response”

Can learn to produce the appropriate response given a particular

task instruction, a context, a reward, etc.

Can either maintain that response or switch to produce a new

one

Can use prediction of reward or outcomes to learn when to use

appropriate task representations